Non-contact identification method for global three-dimensional vibration mode of structure

Abstract

The invention relates to a non-contact recognition method for a global three-dimensional vibration mode of a structure, which comprises the following steps of: acquiring a two-dimensional displacement field of the structure based on a left camera and a right camera in each view field through a full-field optical flow method based on frequency domain characteristics, and calculating according to the two-dimensional displacement field to obtain a sub-pixel displacement field; wherein each view field is subjected to video acquisition by the same binocular camera in a time-sharing manner, and two adjacent view fields have an overlapping region; for each independent view field, projecting a sub-pixel two-dimensional displacement field of the structure to a three-dimensional displacement field according to the parallax between the left image and the right image of the binocular camera, and smoothing the three-dimensional displacement field; converting the three-dimensional displacement field of each view field from the time domain to the frequency domain, and extracting the inherent frequency of the structure and the operation modal shape of each view field under the global coordinate system; and splicing the operation modal shape of each view field based on feature point detection and matching of the overlapping region to obtain a three-dimensional global operation modal shape of the structure. The global vibration mode of the structure can be covered, and the reliability of the recognition result is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of structural health monitoring, and in particular to a non-contact identification method of a global three-dimensional vibration mode of a structure. Background Art

[0002] During operation, civil engineering structures are affected by environmental erosion, material degradation, foundation settlement, and repeated traffic loads, which may cause structural damage and gradual degradation of material performance. In order to monitor the structural condition in real time and obtain various parameter indicators of the structure during service, structural health monitoring (SHM) technology has become an important research direction in the field of civil engineering. This technology uses various advanced sensing technologies to monitor the vibration signals of the structure, and then uses signal processing technology to identify the modal parameters of the structure (such as natural frequency, vibration mode and damping ratio, etc.), and then evaluate the structural performance.

[0003] Traditional vibration measurement methods usually install multiple sensors (such as accelerometers) on the surface of the structure and use the structural response to obtain the natural frequency and vibration mode. However, this method has the disadvantages that the sensors need to be connected to external devices, the number is large, and the installation cost is high. In order to solve these problems, some researchers have proposed non-contact vibration measurement systems, such as laser Doppler vibrometers and radar systems. Although these non-contact measurement devices can achieve high-precision measurements, they are relatively expensive and require special operating skills. In addition, the point characteristics of the above-mentioned vibration sensors limit their high-resolution measurement capabilities, thereby affecting the accuracy of structural dynamic response identification.

[0004] Computer vision-based non-contact measurement systems have become a powerful method for structural vibration measurement due to their significant advantages in instrument cost, workload, and spatial resolution. This technology uses cameras and advanced image processing algorithms to accurately obtain the dynamic response of structures. However, due to the limitations of the camera's field of view, computer vision technology can only measure the vibration response of specific parts of large and complex structures. In order to cover the entire structure, multiple pairs of cameras are usually required for synchronous measurement, but this significantly increases the cost of detection. Therefore, it remains to be developed a low-cost, non-contact method for global three-dimensional vibration mode recognition of complex structures. Summary of the invention

[0005] The present invention aims to solve the deficiencies of the prior art and provide a non-contact identification method for the global three-dimensional vibration mode of a structure.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A non-contact identification method for a global three-dimensional vibration mode of a structure comprises the following steps:

[0008] S1. Stereo calibration of the binocular camera is performed in each separate field of view to obtain a video of the structural vibration; frequency domain features are extracted from the vibration videos of the left and right cameras corresponding to the structure using a two-dimensional filter, and the two-dimensional displacement field of the structure based on the left and right cameras is obtained by a full-field optical flow method based on the frequency domain features of the image, and a sub-pixel displacement field is calculated based on the two-dimensional displacement field;

[0009] Among them, each field of view is captured by the same binocular camera in time-sharing mode and there is an overlapping area between two adjacent fields of view;

[0010] S2, for each individual field of view, projecting the sub-pixel two-dimensional displacement field of the structure into a three-dimensional displacement field according to the parallax between the left and right images of the binocular camera, and smoothing the three-dimensional displacement field;

[0011] S3. Based on the overlapping area of ​​adjacent fields of view, the coordinates of each field of view point are converted into a unified global coordinate system using a three-dimensional translation and rotation matrix; after normalizing the displacement field, the three-dimensional displacement field of each field of view is transformed from the time domain to the frequency domain, and the structural natural frequency and the operating mode vibration shape of each field of view in the global coordinate system are extracted based on the covariance-driven random subspace identification method and hierarchical clustering analysis;

[0012] S4. Based on the feature point detection and matching of the overlapping area, the operating mode vibration shapes of each field of view are spliced ​​to obtain the three-dimensional global operating mode vibration shape of the structure.

[0013] Preferably, in step S1, a two-dimensional Gabor filter is used to extract frequency domain features from the video of structural vibration, and a Gaussian image pyramid iterative algorithm is used to obtain a sub-pixel displacement field.

[0014] Preferably, in step S2, a semi-global block matching algorithm based on image frequency domain information is used to project the two-dimensional displacement field of the structure into a three-dimensional displacement field, and support vector regression is used to smooth the three-dimensional displacement field.

[0015] Preferably, in step S3, the displacement field is normalized by finding the least square solution. Further, the full-field displacement response of the structure can be represented as a Hankel matrix, and then a covariance matrix is ​​constructed based on the Hankel matrix and the covariance matrix is ​​subjected to singular value decomposition to obtain a system matrix, and finally the modal parameters are extracted by eigenvalue decomposition based on the system matrix. At the same time, the stability diagram can be used as a tool to represent the modal parameters, and in the process of automatically selecting the physical mode of the structure from the stability diagram using hierarchical clustering analysis, it includes: treating each modal parameter as a cluster, and merging the two clusters with the smallest inter-cluster distance into a new cluster according to the minimum distance measurement method; recalculating the inter-cluster distance and merging the clusters until the distance between all clusters is greater than the set threshold.

[0016] Preferably, in step S4, feature points of the overlapping area are extracted using a deep learning algorithm based on a self-supervised framework, and the feature points are matched using a deep learning algorithm based on an attention graph neural network and an optimal matching layer.

[0017] The beneficial effects of the present invention are as follows: the present invention can obtain the local vibration modes corresponding to each field of view based on the videos of each field of view collected in time-sharing, and then perform vibration mode splicing based on the overlapping areas in the field of view, so as to cover the global vibration mode of the structure, thereby ensuring the reliability of the recognition result, and is particularly suitable for vibration mode recognition of complex structures; before vibration mode splicing, the displacement fields of each field of view are normalized, and the coordinates of each field of view point are converted into a unified global coordinate system, thereby ensuring the consistency of the scaling scale of vibration mode splicing, and laying a foundation for the subsequent automatic splicing of adjacent vibration modes using feature points; in the process of processing video data, the anti-interference ability is improved through the transformation between the time domain and the frequency domain. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a schematic diagram of the process of the present invention;

[0019] The following is a detailed description of the embodiments of the present invention with reference to the accompanying drawings. DETAILED DESCRIPTION

[0020] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The embodiments are only used to explain the present invention and are not used to limit the scope of the present invention. The present invention is described in more detail by way of example with reference to the accompanying drawings in the following paragraphs. The advantages and features of the present invention will become clearer according to the following description. It should be noted that the accompanying drawings are all in a very simplified form and are not in precise proportions, and are only used to facilitate and clearly assist in explaining the purpose of the embodiments of the present invention.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which the present invention belongs. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more of the related listed items.

[0022] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Specific embodiment 1:

[0024] A non-contact identification method for the global three-dimensional vibration mode of a structure mainly includes four parts: (1) two-dimensional displacement field identification; (2) two-dimensional to three-dimensional projection and displacement optimization; (3) displacement time domain to frequency domain transformation, frequency and vibration mode identification; (4) vibration mode scaling and splicing.

[0025] (1) Two-dimensional displacement field identification:

[0026] The binocular camera is stereo calibrated in each separate field of view to obtain the structural vibration video. The frequency domain features are extracted from the structural image using a two-dimensional Gabor filter. The two-dimensional displacement field of the structure is obtained by the full-field optical flow method based on the frequency domain features of the image. The sub-pixel displacement field is obtained using the Gaussian image pyramid iterative algorithm.

[0027] The full-field optical flow method based on image frequency domain features is as follows:

[0028] Optical flow is based on some assumptions about pixel or feature point matching, such as temporal continuity or small motion, constant brightness, etc.

[0029] Brightness constancy refers to the consistency of intensity information between two adjacent frames of images, which can be expressed in mathematical form as: Where: , The inner edge of the image plane Axis and Axis optical flow vector; , , The pixel grayscale pairs are The partial derivative of are the horizontal and vertical coordinates of the point in the image plane respectively; The timestamp when the required frame image is obtained.

[0030] Preferably, in this embodiment, a Gabor filter may be used to perform a two-dimensional Gabor transform on the image sequence to obtain phase information in a predetermined direction, which is expressed as follows:

[0031]

[0032] Where: represents the input image; is the filtered output image, is the filtering direction; is the convolution symbol; is a two-dimensional Gabor filter, and its function expression is:

[0033] ;

[0034] Where: , Represents spatial frequency in a specific direction; is a two-dimensional Gaussian function; Is an imaginary unit.

[0035] Among them, a single Gabor filter can filter multiple specific directions; in a two-dimensional coordinate system, the number of filtering directions can usually be set to 4, 6 or 8; when 4 directions are used, the direction angles of 0, 90, 180 and 270 are usually selected; when 8 directions are used, the direction angles composed of 45, 135, 225 and 315 can be added; if 6 directions are used, two of 45, 135, 225 and 315 can be selected to form 6 direction angles with the aforementioned 4 direction angles. As a variation, one direction can also be regarded as corresponding to a specific filter.

[0036] After Gabor filtering the image output, it is necessary to Ensure that each direction meets the basic assumptions of optical flow, and after sorting, we can get the expression: .in, The N×2 coefficient matrix is ​​obtained by taking the derivative of x and y for the image phase information in N specific directions obtained by the Gabor filter; is the desired optical flow vector; The N×1 coefficient matrix is ​​obtained by derivation of the image phase information in N specific directions obtained by the Gabor filter with respect to time t. The expression is as follows:

[0037]

[0038] ;

[0039] ;

[0040] in: .

[0041] The core of the optical flow method is to calculate by setting corresponding constraints , At the same time, the Farneback optical flow method uses the grayscale information of two adjacent frames of images to calculate the optical flow information of each pixel in the region of interest. In order to enhance the anti-interference ability of the optical flow method, this embodiment can estimate the optical flow motion of the region of interest based on the phase information (i.e., frequency domain features) obtained by Gabor filtering the two images. The main process is as follows:

[0042] The pixel information of the image is regarded as a two-dimensional function. Assume that a local coordinate system is constructed near the target pixel, and the pixel of interest is located at the center. It is expressed in the local coordinate system as follows:

[0043] ;

[0044] Where: is a two-dimensional column vector, ; is a 2×2 symmetric matrix; is a 2×1 matrix; is a scalar.

[0045] After defining the neighborhood, the neighborhood information is fitted using the weighted least squares method to obtain a six-dimensional coefficient , and then use the coefficient to perform polynomial approximation on the pixel neighborhood, This breaks down into:

[0046] ;

[0047] in, ; ; .

[0048] like is a non-zero matrix, then the displacement between the previous and next frames is It can be expressed as:

[0049] ;

[0050] in: , It is a 2×1 matrix composed of the neighborhood information of the pixels in the previous and next frames.

[0051] In this embodiment, in order to improve the accuracy of optical flow estimation, the Gaussian image pyramid iterative algorithm is integrated into the optical flow algorithm to achieve sub-pixel matching of optical flow. The specific implementation process is as follows:

[0052] Construct an L-layer image pyramid, where and Respectively The initial optical flow estimation value and optical flow calculation result of the layer, the recursive process is:

[0053] ;

[0054] Where: ; and They are respectively the local coordinate system of the tracked point Direction and Direction of the optical flow vector, let the initial optical flow estimate of the top layer =0.

[0055] Measure from the top of the pyramid ( Layer), calculate the optical flow value of the tracked point to the bottom original image for:

[0056] ;

[0057] Where: That is, the sub-pixel displacement of the target pixel between two consecutive frames.

[0058] (2) 2D to 3D projection and displacement optimization:

[0059] In this embodiment, the two-dimensional displacement field of the structure is projected into a three-dimensional displacement field by using a semi-global block matching (SGBM) algorithm based on image frequency domain information; and the three-dimensional displacement field is smoothed by using support vector regression (SVR), as described in detail as follows:

[0060] (2.1) Semi-global block matching algorithm based on image frequency domain information:

[0061] The original SGBM algorithm relies only on image intensity information for matching measurement, which is easily affected by changes in environmental conditions such as weather and lighting. Therefore, in order to achieve accurate and robust measurement of disparity information and reprojection matrix, the image frequency domain information extracted in the (1) two-dimensional displacement field identification step is superimposed and integrated with the original algorithm.

[0062] In order to ensure the stability of image information during SGBM measurement, the grayscale image is first converted into a phase image containing richer image frequency domain information. The results of the Gabor filter in six specific directions are superimposed as the filtering result ,in: , the specific process is as follows:

[0063] The gradient of the image is calculated using the horizontal Sobel operator and Gabor filter results:

[0064] ;

[0065] Map pixels processed by the horizontal Sobel operator to a new image:

[0066] ;

[0067] in Represents the phase information of the point on the new image; Is a constant parameter, the default size is 15.

[0068] Using the phase gradient information of the preprocessed image and the grayscale information of the original image, the gradient cost and the Sum of Absolute Differences (SAD) cost are calculated respectively by a sampling-based method:

[0069] ;

[0070] Where: and are the phase gradient information or intensity information of the left and right images respectively, is the number of pixel samples in the region of interest, The cost of SAD.

[0071] Each pixel in the image is evenly divided into 8 paths, and then constraints are imposed on the one-dimensional paths in eight directions. The global Markov energy equation is derived and energy is accumulated in each direction. All path information is superimposed as the final matching cost of each pixel to obtain the minimum cost path function. , so that approximate two-dimensional constrained matching measurement can be achieved through one-dimensional constraints.

[0072] ;

[0073] Where: is the cost function accumulated along the current path; Indicates the radius or direction of the path; Indicates the position of a pixel, usually using its coordinates ( , )express; represents the disparity value; and are the penalties for smoothness when there is a small difference and a large difference between a pixel and its neighbors, respectively, and < The third term is irrelevant and is only used to eliminate the effect of path lengths in different directions. Adding up the matching costs in the eight directions gives the total matching cost: , and The mathematical expression is:

[0074] ;

[0075] Where: is the number of channels in the image; Indicates the window size of the matching cost SAD (odd number).

[0076] When the lowest cost within the disparity window is (1 + a / 100) times the next lowest cost, the disparity value corresponding to the lowest cost is assigned as the disparity of the pixel. Otherwise, the disparity of the pixel is set to 0. Where a is a constant parameter.

[0077] Sub-pixel interpolation is then performed. The specific interpolation algorithm can refer to many existing technologies and will not be described in detail. After interpolation, assuming that the disparity of a given pixel in the image taken by the left camera is , the parallax of the corresponding pixel captured by the right camera is ,but: ;in, is a constant (usually 1 or 2). If the above formula holds, then Defined as the correct parallax.

[0078] After obtaining the disparity map of the image sequence, measure the disparity change between the previous and next frames Combined with the two-dimensional sub-pixel displacement change obtained in the two-dimensional displacement field recognition step (1), the three-dimensional motion information of each pixel can be calculated as follows:

[0079] ;

[0080] Where: Q is the reprojection matrix; is the scale factor; , , They are , , Direction vibration.

[0081] (2.2) Use SVR to smooth the three-dimensional displacement field:

[0082] Take the displacement of each pixel in the region of interest as the sample x, and let the sample set be , is the number of samples, is the sample dimension. After training, the linear equation in high-dimensional feature space is used To fit the sample set ,therefore It can be described as the following optimization problem:

[0083] ;

[0084] Where: is the inner product of two vectors, For nonlinear mapping, the sample feature space is mapped to a high-dimensional feature space; is the sample penalty factor, is the insensitivity coefficient, and is the size of the deviation sensitive area; and is the model parameter. In the nonlinear case is no longer given explicitly; by constructing the Lagrangian function, it can be determined And the nonlinear regression model:

[0085] ;

[0086] ;

[0087] Where: and is the Lagrange multiplier, is the kernel function.

[0088] (3) Displacement transformation from time domain to frequency domain, frequency and vibration mode identification:

[0089] In the overlapping area of ​​each adjacent field of view, the displacement field is normalized by finding the least squares solution, and the coordinates of each field of view point are converted into a unified global coordinate system using the three-dimensional translation and rotation matrix in the overlapping area. The three-dimensional displacement field of each field of view is transformed from the time domain to the frequency domain, and the structural natural frequency and operating mode vibration shape in the global coordinate system are automatically extracted based on the covariance-driven random subspace identification method and hierarchical clustering analysis.

[0090] (3.1) Find the least squares solution:

[0091] The response of the midpoint in the overlapping area of ​​adjacent fields of view is used for normalization. The frequency domain deformation data of field of view 1 and field of view 2 in the overlapping area are respectively composed of matrices A and B, and the relationship is:

[0092] ;

[0093] In the formula, is the least squares error (mentioned in the above formula and is a 2×2 symmetric matrix. Do they have the same meaning? If not, replace the symbols), multiply both sides of the equation by the transpose of A:

[0094] ;

[0095] Compute the least squares error using matrix inversion:

[0096] ;

[0097] The least squares error obtained is multiplied by the frequency domain deformation data of each point in field of view 2 to scale the operating mode shape.

[0098] (3.2) Extract modal parameters using SSI-COV:

[0099] The full-field displacement response of the structure is expressed as a Hankel matrix, and then the covariance matrix is ​​constructed based on the Hankel matrix and the covariance matrix is ​​subjected to singular value decomposition to obtain the system matrix. Finally, the modal parameters are extracted based on the system matrix through eigenvalue decomposition. The covariance matrix elements are calculated as follows:

[0100] ;

[0101] Where: For the edge point The output vector of sampling intervals, is the number of structural vibration response samples. The Toeplitz matrix (i.e., covariance matrix) in the above processing is as follows:

[0102] .

[0103] (3.3) Automatically select physical modes using hierarchical clustering analysis:

[0104] As a tool for representing modal parameters, the stability diagram requires manual selection by the user to separate physical modes from false modes, and cannot solve the problem of automatic identification of modal parameters. Therefore, hierarchical clustering analysis is used to automatically select the physical modes of the structure from the stability diagram: each modal parameter is regarded as a cluster, and the two clusters with the smallest inter-cluster distance are merged into a new cluster according to the minimum distance measurement method; the inter-cluster distance is recalculated and the clusters are merged until the distance between all clusters is greater than the set threshold. Among them, the distance calculation formula between two modes is as follows:

[0105] ;

[0106] Where: and Respectively Order and order natural frequency; For the Order and Modal assurance criterion between the order mode shapes.

[0107] (4) Mode scaling and splicing:

[0108] Based on the detection and matching of feature points in the overlapping areas, the operating mode shapes of each field of view are spliced ​​to obtain the three-dimensional global operating mode shape of the structure.

[0109] Preferably, this step can utilize SuperPoint, a deep learning algorithm based on a self-supervised framework, to extract feature points of the overlapping area, and match the feature points through SuperGlue, a deep learning algorithm based on an attention graph neural network and an optimal matching layer.

[0110] In this step, the feature points can be selected from corner points, edge points, etc.; wherein each field of view angle corresponds to only a local vibration mode curve, and there is an overlapping area based on adjacent field of view angles, and the vibration mode information calculated based on the overlapping area will exist in the vibration mode curves of the two adjacent field of view angles and have the same characteristics such as trends, so that automatic stitching can be achieved through the detection and matching of feature points.

[0111] In summary, the present invention relates to a non-contact identification method for a global three-dimensional vibration mode of a structure, such as Figure 1 As shown, the core steps include:

[0112] Step S1, stereo calibration of the binocular camera in each separate field of view to obtain a video of the structural vibration; using a two-dimensional filter to extract frequency domain features from the vibration videos of the left and right cameras corresponding to the structure, respectively, and using a full-field optical flow method based on frequency domain features to obtain a two-dimensional displacement field of the structure based on the left and right cameras, respectively, and calculating a sub-pixel displacement field based on the two-dimensional displacement field; wherein each field of view is captured by the same binocular camera in time-sharing and there is an overlapping area between two adjacent fields of view.

[0113] Among them, the calibration of this step is usually to pre-calibrate the internal parameters and distortion coefficients of the left and right cameras of the binocular camera, as well as the relative positions between the two cameras. This is a prior art and will not be described in detail.

[0114] Step S2: for each individual field of view, the sub-pixel two-dimensional displacement field of the structure is projected into a three-dimensional displacement field according to the parallax between the left and right images of the binocular camera, and the three-dimensional displacement field is smoothed.

[0115] In this step, the role of smoothing is mainly to remove sharp burrs, which can be regarded as noise. In vibration type recognition, only the displacement trend needs to be identified. Therefore, smoothing can ensure the reliability of subsequent vibration type recognition and splicing results.

[0116] Step S3, based on the overlapping area of ​​adjacent fields of view, the coordinates of each field of view point are converted into a unified global coordinate system using a three-dimensional translation and rotation matrix; after normalizing the displacement field, the three-dimensional displacement field of each field of view is transformed from the time domain to the frequency domain, and the structural natural frequency and the operating mode vibration shape of each field of view in the global coordinate system are extracted based on the covariance-driven random subspace identification method and hierarchical clustering analysis.

[0117] Step S4: splicing the operating mode shapes of each field of view based on feature point detection and matching in the overlapping area to obtain the three-dimensional global operating mode shape of the structure.

[0118] Inspired by the above core steps, those skilled in the art can flexibly select corresponding algorithms without being limited to the algorithms listed above.

[0119] Similarly, based on the above inspiration, the present invention further develops a non-contact identification system for the global three-dimensional vibration mode of the structure, including a memory and a processor, and implements the above-mentioned non-contact identification method for the global three-dimensional vibration mode of the structure through a computer program stored in the memory and executable on the processor.

[0120] The present invention has the following beneficial effects:

[0121] 1. Since the vibration mode is an inherent property of the structure and is independent of time when the structure does not change, the present invention can obtain the local vibration mode corresponding to each field of view based on the video of each field of view collected in time-sharing, and then perform vibration mode splicing based on the overlapping area in the field of view; thereby covering the global vibration mode of the structure and ensuring the reliability of the recognition result; it is particularly suitable for vibration mode recognition of complex structures.

[0122] 2. Before the vibration mode splicing, the displacement field of each field of view is normalized, and the coordinates of each field of view point are converted into a unified global coordinate system, ensuring the consistency of the scaling scale of the vibration mode splicing, laying the foundation for the subsequent automatic splicing of adjacent vibration modes using feature points.

[0123] 3. In the process of video data processing, the anti-interference ability is improved through the transformation between time domain and frequency domain; for example: the full-field optical flow method based on image frequency domain characteristics can improve the anti-interference ability of displacement extraction.

[0124] The present invention is described above by way of example in conjunction with the accompanying drawings. It is obvious that the specific implementation of the present invention is not limited to the above-mentioned method. As long as various improvements are made using the method concept and technical solution of the present invention, or are directly applied to other occasions without improvement, they are all within the protection scope of the present invention.

Claims

1. A non-contact identification method for global three-dimensional vibration mode of a structure, characterized in that: The steps include: S1. Stereo calibration of the binocular camera is performed in each separate field of view to obtain a video of the structural vibration; frequency domain features are extracted from the vibration videos of the left and right cameras corresponding to the structure using a two-dimensional filter, and the two-dimensional displacement field of the structure based on the left and right cameras is obtained by a full-field optical flow method based on the frequency domain features of the image, and a sub-pixel displacement field is calculated based on the two-dimensional displacement field; Among them, each field of view is captured by the same binocular camera in time-sharing mode and there is an overlapping area between two adjacent fields of view; S2, for each individual field of view, projecting the sub-pixel two-dimensional displacement field of the structure into a three-dimensional displacement field according to the parallax between the left and right images of the binocular camera, and smoothing the three-dimensional displacement field; S3. Based on the overlapping area of ​​adjacent fields of view, the coordinates of each field of view point are converted into a unified global coordinate system using a three-dimensional translation and rotation matrix; after normalizing the displacement field, the three-dimensional displacement field of each field of view is transformed from the time domain to the frequency domain, and the structural natural frequency and the operating mode vibration shape of each field of view in the global coordinate system are extracted based on the covariance-driven random subspace identification method and hierarchical clustering analysis; S4. Based on the feature point detection and matching of the overlapping area, the operating mode vibration shapes of each field of view are spliced ​​to obtain the three-dimensional global operating mode vibration shape of the structure.

2. The non-contact identification method of a global three-dimensional vibration mode of a structure according to claim 1, characterized in that: In step S1, a two-dimensional Gabor filter is used to extract frequency domain features from the video of structural vibration, and a Gaussian image pyramid iterative algorithm is used to obtain a sub-pixel displacement field.

3. The non-contact identification method of a global three-dimensional vibration mode of a structure according to claim 1, characterized in that: In step S2, a semi-global block matching algorithm based on image frequency domain information projects the two-dimensional displacement field of the structure into a three-dimensional displacement field, and support vector regression is used to smooth the three-dimensional displacement field.

4. The non-contact identification method of a global three-dimensional vibration mode of a structure according to claim 1, characterized in that: In step S3, the displacement field is normalized by obtaining the least square solution.

5. The non-contact identification method of a global three-dimensional vibration mode of a structure according to claim 1, characterized in that: In step S4, feature points of the overlapping area are extracted using a deep learning algorithm based on a self-supervised framework, and the feature points are matched using a deep learning algorithm based on an attention graph neural network and an optimal matching layer.

6. A non-contact identification method for a global three-dimensional vibration mode of a structure according to any one of claims 1 to 5, characterized in that: In step S3, the full-field displacement response of the structure is represented as a Hankel matrix, and then a covariance matrix is ​​constructed based on the Hankel matrix and the covariance matrix is ​​subjected to singular value decomposition to obtain a system matrix. Finally, the modal parameters are extracted based on the system matrix through eigenvalue decomposition.

7. A non-contact identification method for a global three-dimensional vibration mode of a structure according to any one of claims 1 to 5, characterized in that: In step S3, the stability diagram is used as a tool to represent the modal parameters, and the physical modes of the structure are automatically selected from the stability diagram using hierarchical clustering analysis, including: treating each modal parameter as a cluster, merging the two clusters with the smallest inter-cluster distance into a new cluster according to the minimum distance measurement method; recalculating the inter-cluster distance and merging the clusters until the distance between all clusters is greater than the set threshold.

Images